Abstract:

To address the two problems of low detector accuracy and redundant keypoints in detection results， a two-dimensional human pose estimation algorithm was proposed based on adaptive parameterized NMS. Replacing the original detector with CenterNet detector improves the performance of human detection and lays the foundation for subsequent pose estimation. Propose an adaptive parameterized PoseNMS algorithm that introduces sample appearance similarity to achieve sample adaptive measurement adjustment， making the filtering conditions in NMS more flexible. A non-uniform sampling method based on detection confidence was proposed， which ensured the effectiveness of the samples during the training process and achieved the discovery and mining of difficult samples. Verified on three datasets， a 71.9% mAP was achieved on the MSCOCO 2017 dataset under the condition of using the detection results output by the detector in this paper as the subsequent target box. In addition， the algorithm proposed in this paper has also been extensively experimented on MPII and MSCOCO 2015， and the quantitative and visual results show that the proposed method effectively solves the above two problems and achieves more accurate attitude estimation.

Key words: CenterNet, key point detection, non-maximum suppression, pose estimation, non-uniform sampling